This is an initial release. This repository currently works well for inference, with no known issues, with training still under development. Current COCO training mAP using this repo is 0.522 (at 416 x 416) after 62 epochs (using all default training settings, simply running python3 train.py). We are exploring ways to improve this further.

Loss curves, Precision, Recall and mAP.

Recommend PyTorch >0 v1.0.0 to run this repo, which includes numerous bug fixes: https://github.com/pytorch/pytorch/releases

Inference

• Inference appears to be working well, no known issues.

Training

• Training is still under development. v1.0 benchmark is currently 0.522 mAP at epoch 62 at 416 image size without multi-scale training.
• Balancing of the various loss terms seem to have great effect on the results, the current constants were tuned for training performance. Further study is needed here.
• Augmentation may be a bit aggressive, experimentation is needed at reduced levels.
• Training seems to suffer jumps in losses (visible in plots above). This may be associated with restarting a stopped training session, the cause is unclear, and some users seem not to experience this.

Validation

• mAP calculation is now correct. mAP is calculated per class per image, and then averaged over images. Despite this, current mAP calculation in test.py seems to be slightly different than the official COCO mAP code.
• It would be very useful for test.py to additionally output text files in the official COCO mAP format.

Multi-GPU

• Multi GPU is currently not supported. PRs welcome!

Performance

• Numerous commits addressed performance issues. Largest changes produced by not reporting P, R, TP, FP, FN each batch, and by reducing data movement between GPU and CPU. Current training speed is about 1 epoch per hour (16-image batches at 416 x 416 each take about 0.7 seconds) on a GCP P100.